Skip to main content

Biren BR100 / BR104 (Domestic AI Training/Inference)

Product Overview​

Biren Technology is a Chinese AI chip startup, founded in September 2019 and headquartered in Shanghai. It listed on the Hong Kong Stock Exchange (HKEX) in January 2025. BR100/BR104 is its first general-purpose GPU chip series, officially released in August 2022.

  • BR100: flagship, dual-chiplet design, 1024 TFLOPS BF16 / 2048 TOPS INT8
  • BR104: single-chiplet version, 32GB HBM2e, 300W TDP, aimed at general-purpose computing

Biren is counted alongside Moore Threads, Jingjia Micro, and Iluvatar CoreX as one of China's "Five Tigers" of AI chip startups, with cumulative funding of $700M+.

Core Specifications​

BR100 (Flagship)​

ParameterValue
ArchitectureBiren (in-house Biren ISA)
ProcessTSMC 7nm
DesignDual chiplet (8 compute dies + 4 HBM2e dies)
BF16 Compute1024 TFLOPS
TF32+ Compute512 TFLOPS
INT8 Compute2048 TOPS
FP32 Compute256 TFLOPS
HBM64GB HBM2e (4 stacks)
Memory Bandwidth2.3 TB/s
Inter-Die Bandwidth800 GB/s (BLink interconnect)
TDP300 W
Release2022-08
StatusDeployed in the Shanghai Intelligent Computing Center's 10,000-card cluster in 2025

BR104 (General-Purpose Version)​

ParameterValue
ArchitectureBiren (in-house Biren ISA)
ProcessTSMC 7nm
DesignSingle chiplet
BF16 Compute512 TFLOPS (about half of the BR100)
TF32+ Compute256 TFLOPS
FP32 Compute128 TFLOPS
INT8 Compute1024 TOPS
HBM32 GB HBM2e
Inter-Die Bandwidth256 GB/s (BLink interconnect)
TDP300 W
Form FactorPCIe Gen4 ×16
VirtualizationSupports up to 4 secure virtual instances
Release2022-08
StatusIn mass production

BR100 vs BR104 Comparison​

MetricBR100BR104
PositioningFlagship trainingGeneral-purpose inference
Chip DesignDual chipletSingle chiplet
BF16 Compute1024 TFLOPS~512 TFLOPS
INT8 Compute2048 TOPS~1024 TOPS
HBM64GB HBM2e32GB HBM2e
Inter-Die Bandwidth800 GB/s256 GB/s
TDP~400W300W
Virtualization8 instances4 instances

Six Key Features of the Biren Architecture​

FeatureDescription
TF32+An improved version of NVIDIA TF32 with higher precision
TDATensor data access accelerator
C-WarpCUDA-like Warp parallel scheduling
BLinkHigh-speed inter-chip interconnect
Unified HBM AddressingMultiple chips share the HBM address space
Secure VirtualizationHardware-level multi-tenant isolation

Vendor Information​

ParameterDetails
CompanyBiren Technology
Founded2019-09
Listing2025-01 HKEX
HeadquartersShanghai
Funding$700M+ (the Series B set a record for a single financing round in China's semiconductor industry)
SoftwareBIRENSUPA (CUDA-like software stack)
CustomersShanghai Intelligent Computing Center, Baidu, ByteDance

Use Cases​

  • ✅ Domestic AI training (BR100, 10,000-card clusters)
  • ✅ Domestic AI inference (BR104, 300W low power)
  • ✅ Government/SOE AI projects (domestic substitution)
  • ❌ CUDA ecosystem lock-in (migration to BIRENSUPA required)
  • ❌ International markets (export controls)